Skip to content

Block API-key credentials and mislead-free model UI in shared deployment #4

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
junwen94 merged 1 commit into from
Sep 29, 2026
Merged

Block API-key credentials and mislead-free model UI in shared deployment #4

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
110 changes: 78 additions & 32 deletions app/src/App.tsx
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -498,6 +498,7 @@ const TRANSLATIONS = {
settings_model_desc: "Qwen3.8-27B runs locally by default — your structures and conversations never leave this machine. Add a cloud API key below only if you want to switch to OpenAI, Claude, or Gemini for a chat.",
settings_model_key_placeholder: "Not set",
settings_model_key_hint: "Stored locally on this machine only — a key is only sent to the provider it belongs to, never anywhere else.",
settings_model_key_unavailable_shared: "Not available in the web app — download the local version to use your own API key.",
settings_databases_heading: "Databases",
settings_databases_desc: "Materials Project needs a free API key to search. Get one at materialsproject.org/api.",
settings_compute_heading: "Compute",
Expand Down Expand Up @@ -566,6 +567,7 @@ const TRANSLATIONS = {
settings_model_desc: "Qwen3.8-27B s'exécute localement par défaut — vos structures et conversations ne quittent jamais cette machine. Ajoutez une clé API cloud ci-dessous uniquement si vous souhaitez utiliser OpenAI, Claude ou Gemini pour une conversation.",
settings_model_key_placeholder: "Non définie",
settings_model_key_hint: "Stockée uniquement en local sur cette machine — une clé n'est envoyée qu'au fournisseur auquel elle appartient, jamais ailleurs.",
settings_model_key_unavailable_shared: "Non disponible dans l'application web — téléchargez la version locale pour utiliser votre propre clé API.",
settings_databases_heading: "Bases de données",
settings_databases_desc: "Materials Project nécessite une clé API gratuite pour la recherche. Obtenez-en une sur materialsproject.org/api.",
settings_compute_heading: "Calcul",
Expand Down Expand Up @@ -634,6 +636,7 @@ const TRANSLATIONS = {
settings_model_desc: "Qwen3.8-27B läuft standardmäßig lokal — Ihre Strukturen und Unterhaltungen verlassen diesen Rechner nie. Fügen Sie unten nur dann einen Cloud-API-Schlüssel hinzu, wenn Sie für einen Chat zu OpenAI, Claude oder Gemini wechseln möchten.",
settings_model_key_placeholder: "Nicht festgelegt",
settings_model_key_hint: "Wird nur lokal auf diesem Rechner gespeichert — ein Schlüssel wird ausschließlich an den zugehörigen Anbieter gesendet, niemals anderswohin.",
settings_model_key_unavailable_shared: "In der Web-App nicht verfügbar — laden Sie die lokale Version herunter, um Ihren eigenen API-Schlüssel zu verwenden.",
settings_databases_heading: "Datenbanken",
settings_databases_desc: "Materials Project benötigt einen kostenlosen API-Schlüssel für die Suche. Holen Sie sich einen unter materialsproject.org/api.",
settings_compute_heading: "Rechenressourcen",
Expand Down Expand Up @@ -702,6 +705,7 @@ const TRANSLATIONS = {
settings_model_desc: "Qwen3.8-27B 默认在本地运行——你的结构和对话内容不会离开这台机器。只有当你想切换到 OpenAI、Claude 或 Gemini 进行对话时，才需要在下方添加云端 API 密钥。",
settings_model_key_placeholder: "未设置",
settings_model_key_hint: "仅保存在本机——密钥只会发送给其所属的服务商，绝不会发往其他任何地方。",
settings_model_key_unavailable_shared: "网页版不支持——如需使用自己的 API 密钥，请下载本地版。",
settings_databases_heading: "数据库",
settings_databases_desc: "搜索 Materials Project 需要一个免费的 API key，可以在 materialsproject.org/api 申请。",
settings_compute_heading: "计算",
Expand Down Expand Up @@ -770,6 +774,7 @@ const TRANSLATIONS = {
settings_model_desc: "Qwen3.8-27B viene eseguito localmente per impostazione predefinita — le tue strutture e conversazioni non lasciano mai questa macchina. Aggiungi una chiave API cloud qui sotto solo se vuoi passare a OpenAI, Claude o Gemini per una chat.",
settings_model_key_placeholder: "Non impostata",
settings_model_key_hint: "Salvata solo localmente su questa macchina — una chiave viene inviata solo al fornitore a cui appartiene, mai altrove.",
settings_model_key_unavailable_shared: "Non disponibile nell'app web — scarica la versione locale per usare la tua chiave API.",
settings_databases_heading: "Banche dati",
settings_databases_desc: "Materials Project richiede una chiave API gratuita per la ricerca. Ottienine una su materialsproject.org/api.",
settings_compute_heading: "Calcolo",
Expand Down Expand Up @@ -1677,7 +1682,13 @@ export default function App() {
const [newProjectChatTitle, setNewProjectChatTitle] = useState("");
const [newProjectWithSources, setNewProjectWithSources] = useState(false);
const [newProjectSourcesText, setNewProjectSourcesText] = useState("");
const [selectedModel, setSelectedModel] = useState(MODEL_GROUPS[0].items[0]);
// Web app is LLM-locked to the local model regardless of selection
// (resolve_model()'s GOLDILOCKS_AGENT_MODEL env-var override always wins
// server-side) -- defaulting to Claude there would misrepresent which
// model actually answers.
const [selectedModel, setSelectedModel] = useState(
SHARED_DEPLOYMENT ? MODEL_GROUPS[1].items[0] : MODEL_GROUPS[0].items[0],
);
// Ephemeral UI-only state (which dropdown is open, is a calc in flight) --
// fine to stay global, same as structure-search's own loading flag/open
// dropdown state. Everything else MLIP-related lives in
Expand Down Expand Up @@ -1917,12 +1928,12 @@ export default function App() {
}, [resolvedTheme]);

useEffect(() => {
refreshConfiguredProviders();
// Projects/history/experience-level are all single-local-user concepts
// with no per-user isolation on the server (design doc §19.2) -- the
// shared deployment's backend 404s these routes anyway, so skip calling
// them at all rather than surfacing that as a startup error.
// Projects/history/experience-level/credentials are all single-local-user
// concepts with no per-user isolation on the server (design doc §19.2) --
// the shared deployment's backend 404s/rejects these routes anyway, so
// skip calling them at all rather than surfacing that as a startup error.
if (!SHARED_DEPLOYMENT) {
refreshConfiguredProviders();
refreshProjects();
loadConversationsOnce();
hydrateExperienceLevel();
Expand Down Expand Up @@ -2591,6 +2602,28 @@ export default function App() {
// (to peek at it, to test something) used to commit whatever was typed the
// moment focus left, silently overwriting a real saved key with no undo.
function renderCredentialRow(id, label) {
if (SHARED_DEPLOYMENT) {
// Server-side, /api/credentials rejects writes here too (server.py) --
// this deployment is LLM-locked to the fixed local model regardless of
// what's typed here (resolve_model()'s env-var override always wins).
// Shown disabled rather than hidden so visitors can see cloud models
// exist and that the web app isn't where you'd use your own key.
return (
<div className="credential-row" key={id}>
<span className="credential-label">{label}</span>
<input
type="password"
className="credential-input"
value=""
placeholder={t("settings_model_key_unavailable_shared")}
disabled
/>
<button type="button" className="secondary-btn compact" disabled>
{t("save")}
</button>
</div>
);
}
const canSave = Boolean(cloudApiKeys[id]?.trim()) && credentialStatus[id] !== "saving";
return (
<div className="credential-row" key={id}>
Expand Down Expand Up @@ -9295,33 +9328,46 @@ export default function App() {
{MODEL_GROUPS.map((group, groupIndex) => (
<div key={group.group}>
{groupIndex > 0 && <div className="sidebar-divider" style={{ margin: "6px 6px 8px" }} />}
<div className="menu-label">{group.group}</div>
{group.items.map((model) => (
<div
key={model.id}
className={`menu-item${model.disabled ? " menu-item-disabled" : ""}`}
aria-disabled={model.disabled || undefined}
onClick={() => {
if (model.disabled) return;
setSelectedModel(model);
setModelOpen(false);
}}
>
<div className="menu-item-icon">
<span className="model-dot" style={{ background: MODEL_TAG_COLORS[model.tag] }} />
</div>
<div className="menu-item-copy">
<strong>{model.label}</strong>
{model.disabled ? (
<span className="model-coming-soon">{model.desc}</span>
) : model.tag !== "default" && configuredProviders[model.tag] ? (
<span className="model-ready">✓ API key saved — ready to use</span>
) : (
<span>{model.desc}</span>
)}
<div className="menu-label">
{group.group}
{SHARED_DEPLOYMENT && groupIndex === 0 ? " — unavailable in web app" : ""}
</div>
{group.items.map((model) => {
// Cloud items stay visible (so visitors know they
// exist) but unselectable here -- same reasoning
// as the Settings credential rows: no safe way to
// bring your own cloud key to a shared server.
const lockedByDeployment = SHARED_DEPLOYMENT && model.tag !== "default";
const isDisabled = model.disabled || lockedByDeployment;
return (
<div
key={model.id}
className={`menu-item${isDisabled ? " menu-item-disabled" : ""}`}
aria-disabled={isDisabled || undefined}
onClick={() => {
if (isDisabled) return;
setSelectedModel(model);
setModelOpen(false);
}}
>
<div className="menu-item-icon">
<span className="model-dot" style={{ background: MODEL_TAG_COLORS[model.tag] }} />
</div>
<div className="menu-item-copy">
<strong>{model.label}</strong>
{lockedByDeployment ? (
<span className="model-coming-soon">Unavailable in the web app</span>
) : model.disabled ? (
<span className="model-coming-soon">{model.desc}</span>
) : model.tag !== "default" && configuredProviders[model.tag] ? (
<span className="model-ready">✓ API key saved — ready to use</span>
) : (
<span>{model.desc}</span>
)}
</div>
</div>
</div>
))}
);
})}
</div>
))}
</div>
Expand Down
12 changes: 12 additions & 0 deletions src/goldilocks_agent/server.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -346,13 +346,25 @@ async def get_chat_history(thread_id: str) -> dict:

@app.post("/api/credentials")
async def save_credential(request: CredentialRequest) -> dict:
# `~/.config/goldilocks/config.toml` is one file on the server, not a
# per-thread/per-user store (unlike the checkpointer) -- letting any
# visitor write to it in the shared deployment would mean the first
# person who opens Settings sets everyone else's provider key (their own
# account gets billed for the whole instance's traffic, silently). This
# deployment mode is LLM-locked to a fixed local model precisely so this
# can't come up (see resolve_model()'s env-var override) -- rejecting
# here too is defense in depth, not just the frontend happening not to
# show the field, same reasoning as `_reject_in_shared_deployment` below
# for projects/conversations.
_reject_in_shared_deployment()
write_credential(request.provider, request.api_key)
return {"ok": True}


@app.get("/api/credentials")
async def get_credentials_status() -> dict[str, bool]:
"""Presence per provider only -- never returns the key itself."""
_reject_in_shared_deployment()
return configured_providers()


Expand Down
16 changes: 16 additions & 0 deletions tests/test_server.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -56,6 +56,22 @@ def test_shared_deployment_blocks_projects_and_conversations_routes(
assert client.delete("/api/conversations/some-id").status_code == 404


def test_shared_deployment_blocks_credentials_routes(tmp_path, monkeypatch) -> None:
"""`~/.config/goldilocks/config.toml` is one file on the server, not a
per-thread/per-user store -- letting any visitor write to it on a shared
server would mean the first person who opens Settings sets everyone
else's provider key (their own account gets billed for the whole
instance's traffic, silently). Refused at the route, not just left
unused by the shared-mode frontend build."""
monkeypatch.setenv("GOLDILOCKS_AGENT_SHARED_DEPLOYMENT", "1")
with _make_client(tmp_path, monkeypatch) as client:
assert client.get("/api/credentials").status_code == 404
save_response = client.post(
"/api/credentials", json={"provider": "anthropic", "api_key": "sk-test"}
)
assert save_response.status_code == 404


def test_shared_deployment_skips_conversation_indexing(tmp_path, monkeypatch) -> None:
monkeypatch.setenv("GOLDILOCKS_AGENT_SHARED_DEPLOYMENT", "1")
calls = []
Expand Down
Loading